ElevenLabs is the best AI audio generator for a company that wants a single platform. It covers voice, music and sound effects, with a commercial licence from around 5.50 € a month. To feed a voice agent, Cartesia costs five times less and answers faster. For music, Suno holds the best quality, but now limits what you are allowed to take away.
This comparison covers eight audio generators, checked in September 2026. Two of them no longer do what other comparisons describe. A third has vanished from the web entirely, and is still recommended everywhere.
How this comparison was made
Criteria come before names, because a ranking whose rules arrive afterwards is worthless. Six questions formed the grid, in this order.
- Price brought back to the unit that matters. One minute of speech for voice, one track for music. Credits exist to hide the gaps, not to explain them.
- The underlying model and where it runs. An audio generator is first a model running somewhere. Knowing where commits your compliance.
- What you can download. The question looked absurd a year ago. It has become the first criterion on music.
- The consent proof required. Cloning a human voice is not a neutral feature in France.
- Whether French is genuinely served. A model trained on English produces an accent your customers hear.
- Whether the tool exists on the day of writing. This criterion should not have been needed. It removed one tool from the list.
The 8 AI audio generators in one table
| Tool | What it produces | Paid entry | Pick it if |
|---|---|---|---|
| ElevenLabs | Voice, music, effects | around 5.50 €/month | you want one single platform |
| Cartesia | Realtime voice | around 4.50 €/month | you are feeding a voice agent |
| OpenAI | Voice via API | usage based, no subscription | you bill by volume |
| Murf | Studio voice-over | around 17.50 €/month | your team does not code |
| Kyutai | Voice, open weights | free, CC-BY licence | you host it yourself |
| Resemble AI | Open source voice and detection | free, usage based | watermarking matters as much as voice |
| Suno | Music | around 7.50 €/month | you produce original music |
| Udio | Music | around 9 €/month | you listen without exporting |
1. ElevenLabs, the most complete of the category
ElevenLabs is the only tool in this list covering all three families of generative audio. Synthetic voice, music and sound effects live behind one subscription and one API. That explains its position more than raw voice quality, which several competitors have now matched.
Since August 2025 the vendor also offers Eleven Music, a music model built on licensing deals signed with Merlin and Kobalt. The first groups around 30,000 independent labels, the second is a major music publisher. Rights holders who opt in earn a royalty, on a split announced as equal between publishing and recording (Music Business Worldwide, August 2025).
Key features
- Synthetic voice in 74 languages with the Eleven v3 model, which accepts intent tags inside the text.
- Instant cloning from the Starter tier, professional cloning from the Creator tier.
- Music with a commercial licence included from the first paid tier.
- Automatic dubbing of an existing video, keeping the original voice.
- Packaged conversational agents, which also makes it a voice agent platform.
- Unused credits roll over for two months, on paid plans only.
Underlying model and limits
ElevenLabs trains its own voice models and depends on no third party there. Hosting is American by default, and contractual commitments such as a DPA stay reserved for the Enterprise tier. The most visible limit in heavy use is economic. Credits burn fast, and a studio producing daily quickly lands on a plan costing several hundred euros a month.
Pricing
Free tier at 10,000 credits a month, with no commercial licence. Starter at around 5.50 € a month for 30,000 credits, roughly half an hour of speech, with commercial licence and instant cloning. Creator at around 20 € for 121,000 credits. Business at around 910 € for six million credits, with low-latency voice announced at 5 US cents a minute (ElevenLabs pricing, September 2026).
What users say
Voice naturalness is the consensus compliment, French included. The recurring complaint is billing readability. The credit system makes it hard to anticipate a project cost before producing it, and several tiers force a large price jump to unlock a single feature.
My take on ElevenLabs
I ranked it as nice to have in my AI tool tier list, not as essential, and I stand by that. It is the reference tool on audio, but it gets expensive as soon as volume rises. We used it to produce personalised welcome videos for every new customer of a platform, with the voice generated by ElevenLabs and the image by an avatar tool. On that use, quality justifies the price because volume stays low. On a voice agent taking hundreds of calls, I move elsewhere.
2. Cartesia, the cheapest realtime voice
Cartesia does not try to be the prettiest voice on the market. It tries to be the fastest and the cheapest, which is exactly what a voice agent needs. On a phone call, an extra tenth of a second of latency is noticed immediately, while one missing shade of intonation is not.
Key features
- Sonic speech synthesis, version 3.6 as of September 2026, built for continuous conversation.
- Ink transcription, letting you take both ends of the chain from one vendor.
- Separate credit and agent-minute billing, useful when you measure cost per call.
- Instant cloning from the plan at around 4.50 € a month.
- Native integration into the main voice agent platforms on the market.
Underlying model and limits
Cartesia develops its own models, Sonic for voice and Ink for transcription. The pricing page documents neither European hosting nor customer-side installation, and its FAQ leaves the self-hosting question visibly unanswered (Cartesia pricing, September 2026). The main limit is qualitative. French voices hold short conversation very well, they tire on a long narrative text.
Pricing
Free tier at 20,000 credits a month, about 27 minutes of synthesis. Pro at around 4.50 € for 100,000 credits, about 133 minutes, with the commercial licence. Startup at around 45 € for 1.25 million credits. Scale at around 275 € for eight million. Per minute, the Pro plan lands around 3.5 euro cents, against roughly 17 cents on the ElevenLabs Creator tier.
What users say
Latency and price come back systematically in positive feedback. The complaint targets the voice catalogue, narrower than the studios offer, and documentation that assumes a developer profile. A non-technical user has no business on this platform.
My take on Cartesia
It is my default recommendation on a voice agent, and I said so in my video tutorial. Cartesia is often the fastest, it is not necessarily the best voice, but it is the cheapest. ElevenLabs has a better voice. On a project where cost per call matters, the trade-off settles quickly. In my field report on voice agents I already recommended a simple stack with Cartesia on synthesis and a small language model for basic conversations. That stack still holds in September 2026, give or take a Sonic version.
3. OpenAI, the price floor of synthetic voice
OpenAI does not sell an audio studio, it sells an API. That is exactly what makes it the cheapest option in this comparison once volume matters. There is no subscription, no expiring credits, no tier to cross to unlock a feature. You pay for what you generate.
Key features
- Eleven preset voices steerable through a style instruction written in plain language.
- No human voice cloning offered, which settles the consent question by absence.
- Realtime models for conversational agents, separate from plain synthesis models.
- Public demo with no account on openai.fm, to judge output before paying.
- Usage billing with no commitment and no monthly quota.
Underlying model and limits
The models are OpenAI's own, served from American infrastructure. Two families coexist and bill differently, which often misleads readers. The tts models are charged per character, the audio models per token. The most binding limit stays the absence of cloning. If your project needs an identified person's voice, this vendor is off topic (OpenAI pricing, September 2026).
Pricing
The tts-1 model is billed at 15 dollars per million characters, roughly 1.2 euro cents for one minute of French speech. The high definition version doubles that. The gpt-4o-mini-tts model comes to 12 dollars per million audio tokens produced. Realtime models are markedly more expensive, around 64 dollars per million output audio tokens, with a lighter version at 20 dollars.
What users say
Value for money is unanimous, and steering by style instruction is widely liked. The most frequent complaint targets the number of voices, judged too small for a brand wanting a recognisable sonic signature. The second targets the French accent, correct but less convincing than at the specialists.
My take on OpenAI
I have not deployed OpenAI speech synthesis as the main building block of a client project, so I stick to what is verifiable. What I do observe on our voice agent stacks is that OpenAI remains the best value on the conversation model side, and that its successive price cuts made automated calling affordable. For voice-over at volume, the per-character grid deserves a serious calculation before signing elsewhere.
4. Murf, the voice-over studio for non-technical teams
Murf solves a problem that APIs do not. A training or marketing team has to produce voice-overs, review them, fix a pronunciation and start again, without going through a developer. Murf gives a full editor for that, with direct integrations into PowerPoint, Google Slides and Canva.
Key features
- More than 200 voices across more than 30 languages and accents, picked in a visual editor.
- Pronunciation control word by word, then applied to every voice in the project.
- Office integrations into PowerPoint, Google Slides and Canva, from the Business tier.
- Automatic translation of projects, reserved for the Enterprise tier.
- Stated compliance with GDPR and SOC 2 on all tiers.
Underlying model and limits
Murf announces an in-house model named Falcon 2 on its homepage banner, presented as a fast synthesis API. The big limit is the counting method. The generation quota is expressed in hours per year, not per month, which reads badly and surprises teams producing in bursts. The Creator tier caps at 24 generation hours a year, about two hours a month (Murf pricing, September 2026).
Pricing
Free tier at 10 minutes of generation, with no download and no commercial rights. Creator at around 17.50 € a month on annual billing, with 24 generation hours a year, unlimited downloads and commercial rights. Business at around 60 € a month for 96 hours a year, with the business licence and office integrations. Enterprise on quote, with single sign-on and a no-training commitment on your data.
What users say
Murf displays a 4.7 rating from the software marketplaces it highlights on its own site. Public feedback praises editor simplicity and production speed compared with a dubbing agency. The complaint targets naturalness, behind the best models on emotional text, and the annual ceiling that blocks production peaks.
My take on Murf
I have not deployed it on a project, so I stick to what documentation and reviews allow me to verify. What I can say is that the buyer profile is clear. A team producing training modules every month and wanting no code will find a coherent tool here. A technical team that can call an API will pay ten times too much for the same result.
5. Kyutai, the only French voice model with open weights
Kyutai is a Paris research lab funded by the Iliad Group, the CMA CGM Group and Schmidt Sciences. It publishes its voice models openly, with the code and the training recipe. It is the only entry in this comparison that does not bill, and the only one whose hosting you choose end to end (Kyutai TTS page).
Key features
- Kyutai TTS 1.6B, an English and French streaming model, released under CC-BY 4.0.
- Kyutai Pocket TTS, 100 million parameters, light enough to run on a CPU in real time.
- Generation latency around 220 milliseconds in a solo setup, per the lab's publication.
- Thirty-two simultaneous streams under 350 milliseconds on a single L40 graphics card.
- Moshi and Unmute, two bricks that let any language model speak and listen.
Underlying model and limits
The architecture rests on the delayed streams modelling framework published by the lab, which starts producing sound before receiving all the text. Weights are distributed on Hugging Face (tts-1.6b-en_fr model). The limits are those of any open model. There is no interface, no contractual support, no availability commitment. You have to hold a graphics card, keep it updated and own the outage. The voice catalogue is also narrower than at a commercial vendor.
Pricing
Zero licence cost. The real cost is your infrastructure. A mid-range graphics card rented from a hosting provider comes to a few tens of euros a month, and absorbs several dozen simultaneous streams. The calculation tips towards Kyutai as soon as you pass a few hours of monthly synthesis.
What users say
The audience for this model is technical, and feedback comes from code repositories rather than software marketplaces. No reliable rating exists on the usual review platforms, and it is better to write that than to pad. What comes back in public discussion is the latency it holds and the quality of its French, surprising for a model that size.
My take on Kyutai
I have not put it into production on a client project, so I stay on what is verifiable. The reason to mention it anyway is simple. It is the only serious answer to the question every legal department asks, which is where the model runs. On a project where voice data must not leave France, none of the seven other options answers, and this one does. If that constraint is yours, a technical scoping before choosing beats a free trial.
6. Resemble AI, the voice cloner turned detector
Resemble AI made its name as a voice cloning platform. In September 2026 its homepage announces that deepfakes are everywhere and sells models to detect them. The Products menu no longer holds a single voice generation entry. The page that sold synthesis at resemble.ai/text-to-speech returns a 404, checked first hand.
Synthesis still exists, but it has changed nature. It now lives under the name Chatterbox, released as open source under the MIT licence, with a multilingual variant covering 23 languages including French (Resemble product page). It is the most interesting move in the category. The clone vendor repositioned on proof, and opened the clone.
Key features
- Chatterbox, an open source synthesis model with zero-shot cloning and emotion intensity control.
- Chatterbox Turbo, 350 million parameters, with breath and laughter markers.
- PerTh watermark applied to produced audio, imperceptible and machine readable.
- Deepfake detection on audio, image and video, billed per second.
- On-premise or air-gapped deployment, offered explicitly.
Underlying model and limits
Chatterbox is an in-house model, distributed with open weights, announced under 200 milliseconds to first sound. The limit is not technical, it is strategic. The company's commercial energy goes to detection, and generation now serves as a lead product. A buyer looking for a voice platform with dedicated support has to factor that in. A buyer looking for a model to install in house gains from it.
Pricing
Chatterbox is free, under the MIT licence. Paid plans cover detection, with a Flex tier billed on usage with no subscription, a Team tier at around 255 € a month and a Business tier at around 735 € a month, both on annual billing. Audio detection comes to 3.5 US cents a second on the free tier, and 1.5 cents on paid tiers.
What users say
Published reviews mostly cover the former cloning offer, which makes them of little use for judging the current product. No recent, representative rating exists on the detection side, and it is better to say so than to dress up a number. What is verifiable is the official Microsoft Teams integration the vendor highlights.
My take on Resemble AI
I have not used it on a project, neither for voice nor for detection, so I stay with the observation. That observation alone is worth a read. When the vendor who sold the best cloning starts selling clone detection, it says something about the market. Value moved from making to proving, and that should weigh on your vendor choice.
7. Suno, the best generative music, with conditions
Suno produces the most convincing generative music on the market, structure and vocals included. In November 2025 the vendor settled its dispute with Warner Music Group and signed the first licensing deal between a major and a music generator (Music Business Worldwide, November 2025). That deal changed the product as much as the law.
Key features
- Full track generation with lyrics, sung vocals and structure, from a written prompt.
- Stem separation, two types on Pro, three on Premier.
- Suno Studio, an editing environment reserved for the Premier tier.
- Audio upload up to thirty minutes, to rework an existing recording.
- Commercial licence included on paid plans, absent from the free plan.
Underlying model and limits
Suno trains its own models and does not publish their architecture. The decisive limit is no longer sound quality, it is contractual. Credits included in a subscription roll over neither from day to day nor month to month. Above all, the number of tracks you can download is now disconnected from the number you can generate.
Pricing
Free tier at 50 credits a day, with no download at all and no commercial rights. Pro at around 7.50 € a month for 2,500 credits, roughly 500 generatable tracks, and 20 monthly downloads. Premier at around 22 € for 10,000 credits and 60 downloads. Separately purchased credits do not expire but require an active subscription (Suno pricing, September 2026).
What users say
Musical quality impresses, musicians included, and that is the dominant compliment. The complaint changed nature in a year. It used to target sound artefacts, it now targets download quotas and legal uncertainty around tracks produced before the licensing deals.
My take on Suno
I ranked it among the merely decent tools, noting it is fast and good for mock-ups. That is still my view. To lay an atmosphere on a demo video or test a sonic identity, it does the job in minutes. For a client deliverable released publicly, the download ceiling and the haze over prior rights call for a real reading of the terms before committing.
8. Udio, the licensed platform nothing leaves
Udio was long Suno's direct competitor on quality. On 29 October 2025 the company settled its dispute with Universal Music Group and announced a licensed creation platform (Billboard, October 2025). The vendor's help centre spells out what that deal changed for subscribers, in an article dated 18 February 2026.
Key features
- Track generation from a prompt, with extension and remixing.
- Raised credit ceilings since the deal, to 2,400 a month on Standard and 6,000 on Pro.
- Five simultaneous sets on the Pro tier, meaning ten tracks in parallel.
- Mobile app available on iOS.
- Listening and sharing on the platform, with no file export.
Underlying model and limits
Current models are still those from before the deal. The licensed platform announced for 2026 is to rest on a new generation trained on a cleared catalogue. The limit is head-on and fits in one sentence. You cannot take your production off the platform (Udio help centre, February 2026). Warner Music also settled with Udio, while Sony Music was still in proceedings at the time of writing.
Pricing
Free tier at 100 credits a month. Standard at around 9 € a month for 2,400 credits, Pro at around 28 € for 6,000 credits, with 20 percent off on annual billing. One-off top-ups exist, and they do not expire. These prices buy generation capacity, not an export right.
What users say
The musical output keeps convinced defenders, notably on acoustic genres. The single topic of recent feedback stays the download shutdown, experienced as a unilateral rule change by subscribers who had built a catalogue. That is the dominant complaint, and it eclipses everything else.
My take on Udio
I do not use it on projects, and the platform's current state makes that question secondary. For a company, a tool whose output does not leave is not a tool, it is a toy. The situation may change when the licensed platform ships. Until it ships, this vendor does not go into a production chain.
Which AI audio generator to choose for your use
A comparison that does not rank is a directory. Here is the recommendation by profile, with the reason behind it.
- You want one platform for voice and music. ElevenLabs. It is the only one covering both with a commercial licence from the first paid tier.
- You feed a voice agent at volume. Cartesia. Five times cheaper than ElevenLabs per minute, with latency built for the phone.
- You produce voice-over at scale from code. OpenAI. Around 1.2 cents a minute, with no subscription and no expiring credits.
- Your marketing or training team does not code. Murf. The visual editor and office integrations justify the premium, provided you read the annual quota.
- Your voice data must not leave France. Kyutai. It is the only complete answer, at the price of infrastructure you run.
- You need to prove an audio file is authentic. Resemble AI, for watermarking and detection, with Chatterbox for self-hosted generation.
- You want original, usable music. Suno, after computing your real download need before picking a tier.
- You need music you can export today. Neither free Suno nor Udio. Google's Lyria API bills around 7 cents a track, downloadable.
Three things audio generator comparisons do not say
French criminal law has punished voice cloning since 2024, before the AI Act
The European AI regulation applies since 2 August 2026, and its article 50 imposes two distinct obligations many people confuse. The provider must mark synthetic content in a machine-readable format. Whoever distributes must perceivably disclose that the content is a deepfake (EU regulation 2024/1689). The vendor's invisible mark therefore releases you from nothing.
But French criminal law got there first. Article 226-8 of the criminal code, amended by the law of 21 May 2024, punishes by one year of imprisonment and a 15,000 € fine the distribution of a montage made with a person's words without their consent, where it is not obvious that it is a montage. Penalties rise to two years and 45,000 € when distribution runs through an online service (article 226-8, Légifrance). That text targets voice explicitly, and it applies whether the tool is American or not.
This is not a theoretical risk. In February 2026, eight French dubbing actors sent formal notices to two American AI companies, Voice Dub and Fish Audio, for cloning their voices without agreement. Brigitte Lecordier, the French voice of Son Goku, asks for traceability on voices and their removal from the platforms. Dubbing represents around one billion euros of annual revenue and 15,000 jobs in France (franceinfo, February 2026). The campaign is carried by the TouchePasMaVF collective and the French performers' union.
On music, the entry price no longer says anything about real cost
Generating and taking away have become two separate operations, billed separately. The Suno Pro plan at around 7.50 € covers roughly 500 tracks and allows 20 downloads. At Udio, downloading is disabled whatever you pay. On the French side, Sacem exercised its opt-out from text and data mining across its whole repertoire as early as October 2023, on the basis of article L.122-5-3 of the intellectual property code (Légipresse, October 2023). Any use of that repertoire by an AI developer therefore requires prior authorisation.
Practical consequence for a company. Count in deliverable tracks, never in generated tracks, and keep a record of the plan each file was produced under. One alternative exists and nobody cites it, Google's Lyria API bills around 7 euro cents a track, with no subscription and no export ceiling (Gemini API pricing, September 2026).
The market is emptying from the top, and comparisons do not follow
PlayHT was still in the reference list for this category when this comparison was prepared. The service no longer exists. Meta hired the team in July 2025 (TechCrunch, July 2025), the platform closed at the end of 2025, and the play.ht domain simply no longer resolves in DNS, checked first hand in September 2026. Resemble AI, for its part, removed voice generation from its navigation to sell detection.
The lesson is useful beyond audio. On categories consolidating fast, the first check is not price comparison, it is product existence. It takes two minutes and it removed one name out of nine from this list.
Which AI audio generator to pick, one sentence per use
ElevenLabs if you want one platform and volume stays reasonable. Cartesia as soon as you industrialise realtime voice, because the price gap reaches a factor of five. OpenAI if you can call an API and voice-over ships at high volume. Kyutai if your constraint is hosting, and then it is the only answer.
On music, the trade-off moved ground. The question is no longer which one sounds best, but which one lets you leave with your files. Suno on a paid plan sized in downloads, or the Lyria API if you accept going through code. Udio will wait for its licensed platform.
What these eight tools share is that the best choice depends on your volume and your legal constraint, not on the beauty of the demo. That is exactly what we look at first on an automation scoping, and it is also what separates a project that holds from a subscription that sleeps. If the subject is the voice agent rather than the voice itself, platform choice is handled in our AI voice agent comparison, and deployment method in our voice agent guide.
